Re: JPEG-LS call for proposals?

Sachin Garg wrote:

I read in a paper that the call for proposals specified that the
algorithms should *not* use inter-component correlations (which can
improve ratios). While there are some advantages of keeping the
components separate in some cases, I was wondering why exactly it was
so important.

The same requirement was formulated for JPEG2000, but the reason here
was that not using inter-component correlations allow you to use various
progression modes, i.e. alter the order in which components/spatial image regions/resolutions arrive. As soon as you make use of inter-component correlations, you need to ensure that the components that participate in the decorrelation transformation arrive in a sensible order so the decoder can perform its job. Thus flexibility
was rated higher than the advantage in compression performance.

For JPEG-LS, the algorithm was always considered as "codestream only"
and any kind of color-decorrelation has to be done by an external
file-format wrapper that maps the original color space into a suitably
decorrelated one. For example, ICC profiles could indicate that the
encoded color space is YCbCr, giving the correlation transformation
from YCbCr to XYZ in the profile, and thus the wrapper. Thus, the
color decorrelation is, as far as JPEG-LS is concerned, a "somebody
else's problem"(tm).

In JPEG2000, the color space problem is approximately attacked in the same way (a "SEP"), but the standard at least allows you to use a multi-
component decorrelation as an "afterburner" that, by "pure chance", coincides with an RGB to YCbCr transformation. In the understanding of
the standard, this is however not a color transformation, but a transformation for decorrelating the components. Color space information
has again to go into the file format, and an external color management
module in the sense of ICC has to take care of this.

Many of these points probably would have been same for Jpeg 2000, maybe
you can try looking for that one (but even this will need to go way
back to 1996/1997).

Well, see above. Did that help?

So long,
Thomas
.